#ifndef _INLINE_SSE_H_ #define _INLINE_SSE_H_ #include "sse_mult_nn.h" #include "sse_mult_na.h" #include "sse_mult_an.h" #include "sse_mat_vec.h" #include "sse_adj_mat_vec.h" #include "sse_mat_vec_sum_4dir.h" #include "sse_adj_mat_vec_4dir.h" #include "sse_adj_mat_vec_4vec.h" #include "sse_su3_proj.h" #include "sse_mat_hwvec.h" #include "sse_adj_mat_hwvec.h" #include "sse_sub4vecs.h" #include "sse_addvec.h" #include "sse_s_m_a_vec.h" #include "sse_s_m_a_mat.h" #ifdef SSE_SUBS #define mult_su3_nn(args...) _inline_sse_mult_su3_nn(## args) #define mult_su3_na(args...) _inline_sse_mult_su3_na(## args) #define mult_su3_an(args...) _inline_sse_mult_su3_an(## args) #define mult_su3_mat_vec(args...) _inline_sse_mult_su3_mat_vec(## args) #define mult_adj_su3_mat_vec(args...) _inline_sse_mult_adj_su3_mat_vec(##args) #define mult_su3_mat_vec_sum_4dir(args...) _inline_sse_mult_su3_mat_vec_sum_4dir(##args) #define mult_adj_su3_mat_vec_4dir(args...) _inline_sse_mult_adj_su3_mat_vec_4dir(##args) #define mult_adj_su3_mat_4vec(args...) _inline_sse_mult_adj_su3_mat_4vec(##args) #define su3_projector(args...) _inline_sse_su3_projector(##args) #define mult_su3_mat_hwvec(args...) _inline_sse_mult_su3_mat_hwvec(##args) #define mult_adj_su3_mat_hwvec(args...) _inline_sse_mult_adj_su3_mat_hwvec(##args) #define sub_four_su3_vecs(args...) _inline_sse_sub_four_su3_vecs(##args) #define add_su3_vector(args...) _inline_sse_add_su3_vector(##args) #define scalar_mult_add_su3_vector(args...) _inline_sse_scalar_mult_add_su3_vector(##args) #define scalar_mult_add_su3_matrix(args...) _inline_sse_scalar_mult_add_su3_matrix(##args) #endif typedef struct { unsigned int c1,c2,c3,c4; } sse_mask __attribute__ ((aligned (16))); static sse_mask _sse_sgn13 __attribute__ ((unused)) ={0x80000000, 0x00000000, 0x80000000, 0x00000000}; static sse_mask _sse_sgn24 __attribute__ ((unused)) ={0x00000000, 0x80000000, 0x00000000, 0x80000000}; static sse_mask _sse_sgn3 __attribute__ ((unused)) ={0x00000000, 0x00000000, 0x80000000, 0x00000000}; static sse_mask _sse_sgn4 __attribute__ ((unused)) ={0x00000000, 0x00000000, 0x00000000, 0x80000000}; #endif